"AI bookkeeping" gets thrown around loosely. Here's a grounded look at what AI actually does well in bookkeeping today, what it doesn't, and how to tell real automation from marketing.
What AI genuinely automates
Transaction categorization
This is the biggest win. AI reads each transaction and predicts the right category with a confidence score, learning from your corrections. Over time, most transactions categorize themselves and you review only low-confidence exceptions.
Anomaly detection
A nightly AI sweep can flag duplicates, unusual spikes, likely miscategorizations, and missing data — turning error-hunting into a short, prioritized list instead of a month-end scramble.
Close assistance
AI can advise during the monthly close — surfacing what looks off before you sign off — while deterministic checks handle the things that must be exactly right.
What still needs a human
- Judgment calls on genuinely ambiguous transactions.
- Tax strategy and unusual structuring decisions.
- Relationship and advisory work.
- Final sign-off on the books.
How to evaluate AI bookkeeping tools
- Does it show confidence scores, or just guess silently?
- Does it learn your rules, or ask about the same vendor forever?
- Does it detect anomalies proactively?
- Can you keep your own ledger (QuickBooks/Xero) or export freely — no lock-in?